Table 9.

The identified best models for all benchmark datasets using resPsoCnn. TB indicates a transitional block which contains a single 1 × 1 convolutional layer and RB indicates a ResNet block which contains two convolutions, as indicated in Figure 3. FC indicates a fully connected layer.

Dataset	Structure
MNIST [3,52]	TB($c_{in}=1$ $c_{out}=176$) + RB(176 × 4 × 4) + RB(176 × 5 × 5) + RB(176 × 5 × 5) + RB(176 × 4 × 4) + RB(176 × 5 × 5) + RB(176 × 3 × 3) + AveragePool + TB($c_{in}=176$ $c_{out}=198$) + RB(198 × 5 × 5) + RB(198 × 6 × 6) + RB(198 × 4 × 4) + RB(198 × 4 × 4) + AveragePool + FC
MNIST-RD [53,54]	TB($c_{in}=1$ $c_{out}=184$) + RB(184 × 4 × 4) + RB(184 × 4 × 4) + RB(184 × 5 × 5) + RB(184 × 4 × 4) + RB(184 × 3 × 3) + RB(184 × 3 × 3) + AveragePool + TB($c_{in}=184$ $c_{out}=146$) + RB(146 × 4 × 4) + RB(146 × 4 × 4) + RB(146 × 5 × 5) + RB(146 × 3 × 3) + AveragePool + FC
MNIST-RB [53,54]	TB($c_{in}=1$ $c_{out}=216$) + RB(216 × 5 × 5) + RB(216 × 6 × 6) + AveragePool + TB($c_{in}=216$ $c_{out}=158$) + RB(158 × 7 × 7) + RB(158 × 4 × 4) + Ma × Pool + FC
MNIST-BI [53,54]	TB($c_{in}=1$ $c_{out}=188$) + RB(188 × 4 × 4) + RB(188 × 5 × 5) + RB(188 × 5 × 5) + RB(188 × 4 × 4) + RB(188 × 4 × 4) + RB(188 × 3 × 3) + RB(188 × 3 × 3) + AveragePool + TB($c_{in}=188$ $c_{out}=177$) + RB(177 × 5 × 5) + RB(177 × 3 × 3) + RB(177 × 3 × 3) + RB(177 × 3 × 3) + Ma × Pool + FC
MNIST-RD+BI [53,54]	TB($c_{in}=1$ $c_{out}=231$) + RB(231 × 4 × 4) + RB(231 × 5 × 5) + RB(231 × 5 × 5) + RB(231 × 4 × 4) + RB(231 × 3 × 3) + RB(231 × 4 × 4) + AveragePool + TB($c_{in}=231$ $c_{out}=120$) + RB(120 × 5 × 5) + RB(120 × 5 × 5) + RB(120 × 3 × 3) + RB(120 × 4 × 4) + RB(120 × 4 × 4) + RB(120 × 3 × 3) + RB(120 × 3 × 3) + AveragePool + FC
RECTANGLES-I [53,54]	TB($c_{in}=1$ $c_{out}=71$) + RB(71 × 4 × 4) + RB(71 × 3 × 3) + RB(71 × 7 × 7) + RB(71 × 6 × 6) + RB(71 × 6 × 6) + RB(71 × 6 × 6) + TB($c_{in}=71$ $c_{out}=21$) + RB(21 × 5 × 5) + RB(21 × 4 × 4) + RB(21 × 6 × 6) + RB(21 × 6 × 6) + RB(21 × 7 × 7) + RB(21 × 4 × 4) + FC